基于GAN与分离卷积的高解码精度图像隐写  

High-Decoding Accuracy Image Steganography Based on GAN and Separated Convolution

作　　者：林俏伶 刘昌华 刘沛 LIN Qiaoling;LIU Changhua;LIU Pei(School of Mathematics and Computer,Wuhan Polytechnic University,Wuhan 430023,China)

机构地区：[1]武汉轻工大学数学与计算机学院,湖北武汉430023

出　　处：《软件导刊》2024年第10期179-186,共8页Software Guide

基　　金：湖北省高等学校省级教学研究项目(2022343)。

摘　　要：从提高秘密图像安全性、解码精度和缩短编解码时间3个方面进行考虑,提出一个基于GAN和分离卷积的高解码精度图像隐写方案。在嵌入秘密信息前使用基于Residual-Rep结构、Inception-SCS结构的预处理网络自动学习载体图像高维特征并以数据驱动方式使用特征表示,全面获取通道和空间的特征信息,降低图像失真,引入残差连接,降低秘密信息损失,通过缩短编码和解码时间降低模型复杂度。在基于稠密结构的解码网络引入了纠错层、纠错函数和沃瑟斯坦度量,提升秘密信息恢复精度。在典型环境下,获得了平均0.89的解码精度,结构相似性平均为0.95,在提高解码精度的同时也降低了图像失真,编码时间比SteganoGAN、Hidden方法均缩短一半,可在更短时间内完成编码任务。The problem of low decoding accuracy and image visual quality,long encoding and decoding time are in existing image steganography.In view of the above challenges,a high-decoding accuracy image steganography based on GAN and separated convolution is proposed.An Residual-Rep structure-based and Inception-SCS structure-based preprocessing network is used to automatically learn the high-dimensional features of the image and use the feature representation in a data-driven way before embedding the secret information,acquiring feature information for both channels and spaces,and the skipping connection is used to reduce the loss of secret information,and reduce model complexity by shortening encoding and decoding time.In order to improve the dense decoder’s accuracy,the error correction layer,error correction function and Wasserstein distance are introduced.In a typical environment,an average decoding accuracy of 0.89 and an average structural similarity of 0.95 are obtained,which improves the decoding accuracy and reduces image distortion.The encoding time is reduced by half compared to both SteganoGAN and Hidden methods,allowing the encoding task to be completed in a shorter time.

关 键 词：生成式对抗网络 沃瑟斯坦度量 残差连接 分离卷积 纠错层 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]